Captions and translations

Transcription gets names wrong. The caption panel on the episode page takes the word as heard and the word as it should be, and corrects every occurrence in the episode at once — because "you heard this name wrong throughout" is the actual complaint, not "fix instance 14".

Corrections are stored per word instance with the original kept alongside, so each one is verified at render time rather than assumed. If a re-transcribe moves or changes that word, the edit is skipped and logged instead of rewriting an unrelated word.

The fix is applied at the single point words are fetched for rendering, which is why the burned captions, the SRT and the VTT cannot disagree with each other.

Every finished clip carries captions.srt and captions.vtt, both on the clip's own timeline rather than the source's — so they stay in sync even when silence has been cut out of the middle.

On any ready clip, open Translated captions, choose a language and press Translate. You get an SRT and a VTT for that language, and the player gains a subtitle track for it.

Cue timings are copied, never re-derived. The translation is built from the cue list of the clip's own VTT, and the model is never shown and never asked for a timestamp — so the only thing it can get wrong is the words. A reply with the wrong number of cues is refused outright rather than realigned, because dropping one cue would shift every later line onto the wrong moment.

Available: Spanish, Portuguese (Brazil), French, German, Italian, Dutch, Polish, Hindi, Japanese, Korean, Chinese (Simplified), Arabic.

Burned-in captions stay in the original language — a second burn is a second render.